diff --git a/ggml/src/ggml-qnn/graph.hpp b/ggml/src/ggml-qnn/graph.hpp
index 9941365f7..c82b7d66a 100644
--- a/ggml/src/ggml-qnn/graph.hpp
+++ b/ggml/src/ggml-qnn/graph.hpp
@@ -86,7 +86,7 @@ public:
             return false;
         }
 
-        QNN_LOG_DEBUG("graph name %s, add_nodes start", _graph_name.c_str());
+        QNN_LOG_DEBUG("graph name %s, build_graph start", _graph_name.c_str());
         _qnn_tensor_inputs.resize(tensor_inputs.size());
         _tensor_inputs.resize(tensor_inputs.size());
         for (size_t i = 0; i < tensor_inputs.size(); i++) {
@@ -111,7 +111,7 @@ public:
             snprintf(buffer, GGML_MAX_NAME, "dst%d", (int)i);
             auto qnn_tensor =
                 std::make_shared<ggml_qnn_tensor>(std::string(buffer), _device, _graph_handle, _qnn_instance);
-            auto *ggml_tensor = tensor_inputs[i];
+            auto *ggml_tensor = tensor_outputs[i];
             if (!qnn_tensor->bind_ggml_tensor(ggml_tensor, false)) {
                 QNN_LOG_ERROR("bind tensor %s failed\n", ggml_get_name(ggml_tensor));
                 return false;
@@ -155,7 +155,7 @@ public:
             return false;
         }
 
-        QNN_LOG_DEBUG("graph name %s, add_nodes succeed", _graph_name.c_str());
+        QNN_LOG_DEBUG("graph name %s, build_graph succeed", _graph_name.c_str());
         return true;
     }
 
@@ -173,7 +173,7 @@ public:
         }
 
         for (size_t i = 0; i < tensor_outputs.size(); i++) {
-            auto *ggml_tensor = tensor_inputs[i];
+            auto *ggml_tensor = tensor_outputs[i];
             if (!_tensor_outputs[i]->bind_ggml_tensor(ggml_tensor, false)) {
                 QNN_LOG_ERROR("bind tensor %s failed\n", ggml_get_name(ggml_tensor));
                 return false;
@@ -191,6 +191,14 @@ public:
             }
         }
 
+        for (auto tensor : _tensor_inputs) {
+            tensor->unbind_ggml_tensor();
+        }
+
+        for (auto tensor : _tensor_outputs) {
+            tensor->unbind_ggml_tensor();
+        }
+
         if (error != QNN_SUCCESS) {
             QNN_LOG_INFO("error = %d\n", error);
             return false;
diff --git a/ggml/src/ggml-qnn/tensor.hpp b/ggml/src/ggml-qnn/tensor.hpp
index 5e45266b4..7709936ed 100644
--- a/ggml/src/ggml-qnn/tensor.hpp
+++ b/ggml/src/ggml-qnn/tensor.hpp
@@ -28,17 +28,22 @@ public:
         QNN_LOG_DEBUG("create tensor %s, device: %d", _tensor_name.c_str(), device);
     }
 
+    ~ggml_qnn_tensor() {
+        if (_qnn_instance && _qnn_rpc_buffer) {
+            _qnn_instance->free_rpcmem(_qnn_rpc_buffer);
+        }
+    }
+
     bool bind_ggml_tensor(ggml_tensor *tensor, bool is_input) {
         if (_tensor) {
             if (_tensor != tensor) {
                 QNN_LOG_WARN("tensor %s has been bound to another ggml tensor %s", _tensor_name.c_str(),
                              ggml_get_name(_tensor));
                 return false;
-            } else {
-                QNN_LOG_INFO("tensor %s already bound to same ggml tensor %s", _tensor_name.c_str(),
-                             ggml_get_name(_tensor));
-                return true;
             }
+            QNN_LOG_INFO("tensor %s already bound to same ggml tensor %s", _tensor_name.c_str(),
+                         ggml_get_name(_tensor));
+            return true;
         }
 
         update_params_from_ggml_tensor(tensor);
@@ -55,13 +60,16 @@ public:
                 return false;
             }
             QNN_TENSOR_SET_ID(_qnn_tensor, QNN_TENSOR_GET_ID(qnn_tensor));
+            QNN_LOG_DEBUG("create graph tensor %s, id: %d", _tensor_name.c_str(), QNN_TENSOR_GET_ID(qnn_tensor));
         }
 
         if (should_use_mem_handle()) {
-            _qnn_rpc_buffer = alloc_rpc_mem(ggml_nbytes(tensor));
             if (!_qnn_rpc_buffer) {
-                QNN_LOG_WARN("alloc rpc mem failed, tensor %s", _tensor_name.c_str());
-                return false;
+                _qnn_rpc_buffer = alloc_rpc_mem(ggml_nbytes(tensor));
+                if (!_qnn_rpc_buffer) {
+                    QNN_LOG_WARN("alloc rpc mem failed, tensor %s", _tensor_name.c_str());
+                    return false;
+                }
             }
 
             QNN_LOG_DEBUG("tensor %s, use mem handle %p", _tensor_name.c_str(), QNN_TENSOR_GET_MEM_HANDLE(_qnn_tensor));
@@ -107,8 +115,8 @@ public:
             QNN_LOG_DEBUG("tensor %s, clear client buffer", _tensor_name.c_str());
         }
 
+        QNN_LOG_DEBUG("unbind tensor: %s from ggml tensor: %s", _tensor_name.c_str(), ggml_get_name(_tensor));
         _tensor = nullptr;
-        QNN_LOG_DEBUG("unbind tensor: %s from ggml tensor: %s", _tensor_name.c_str(), _tensor->name);
         return true;
     }